ComfyUI - Live Stream! Let's make some amazing art with Stable Diffusion!

[Music] pap [Music] la [Music] a [Music] [Music] la hey everybody Welcome to the last Saturday of 2023 and our live stream today uh comfy UI Ai and all kinds of Art and stuff in general uh so today we're going to do some uh some fun stuff we're going to do we're going to take a photo that I did and it's kind of the backside of a ballerina at an event and we're going to change her dress we're going to try and use like a practical application of AI and basically how would we problem solve through this thing so it's kind of a dress designers nightmare to come up with really unique ideas and what I want to do is I want to show some ways to do that uh so this is obviously the last live stream of this year and obviously a big year it was we went from what not even two years ago this wasn't even a thing uh to this whole last year has just been amazing fire hose of technology to what will the next year hold who knows uh we do have some exciting stuff coming out soon like I'll be talking more about the API uh from stability where we now have a video API as you can call that and uh text to video uh so that's obviously uh groundbreaking and some of the stuff that we're doing with that and where we're going with it is just fantastic so a lot of stuff coming along that way feel free to offer some stuff in the chat here prompt wise again think dress designs things along that lines and I'll show you kind of the workflow and we're going to start from the beginning and work our way through now we're going to start with a basic sdxl layout and we're going to use sdxl we might use the Juggernaut model today just cuz it's different and I like it it's very cool there's a lot of really nice fine tunes that have been done to the sdxl base uh so that's kind of what we're going to play with today uh this is this this channel has never been more about hey let's just talk about new models that are coming out and things along that lines there are a few times when I've been very excited about them um but maybe I'll do a little bit more of that as as different shows and talk about hey this is the model that I think is really interesting this week I don't know you tell me what you want to talk about or what you want to hear uh so a couple other things too uh thanks to all the people who supporting this program so obviously a big list at the beginning there uh we are one member shy of 300 so that's a huge number uh so thank you guys so much for keeping this thing going and your support actually makes this show happen another one uh today uh we have a sponsor today actually uh gigabyte sent me a an a 17x laptop to do AI inference and we used it last live stream we're going to use it again this week uh so I have pretty much abandoned my MSI laptop which was only like two months old uh for this one and they're pretty much the same model the same chipset same memory actually this has 2 GB more of video RAM in it than that one did actually 4 gby more um but this laptop just feels so much nicer so I'm talk about that a little bit I'm not going to make it a big ad thing but it is the laptop we're going to be using for inference today so I'm going to be showing this screen here we have my uh local machine and then we're going to switch to the laptop's local machine where it's going to be doing all the heavy lifting and we're going to be using this machine for the UI U that I'm doing that basically so that the live stream doesn't trip up as we're working with it we don't have lag occur because we're trying to do live streaming at the same time that we're doing this uh what's my regular live stream schedule it's supposed to be Saturdays at 10:00 uh I did not do last week obviously because it was Christmas Christmas time and at family in town and my wife would have shot me so best to just kind of skip it and live to see another day so obviously getting back into it so it' be uh pretty much every Saturday now I do have a lot coming up in January February for speaking engagement so I'll be in Vermont I'll be in California I'll be in Texas I'll be in got one other one in there coming up Orlando so I have all these other ones that this is for the photography associations and uh using AI in the photography business which is kind of why I wanted to do this one today is to show more about practical application of how to use this to problem solve now that's important because if you're looking to get into AI as something is a full-time job you're not going to find that that a lot of people are looking for you know people who can prompt engineer wayu models because that seems to be the thing everybody's into I see all these these posts on Reddit it's like hey what do you think of my new AI girlfriend I'm like I'm not even going to comment on I'm just going to I'm just going to Breeze right on by that one but there are a lot of practical applications that can be done for AI in Industry especially design so architecture dress designs fabric designs patterns how do we do that what do we do with that uh so today what I going to do is talk about kind of problem solving that and I'm just I haven't done this yet so this is kind of like I I want to challenge myself to just kind of do it and I like these live streams for that because it feels more genuine like we're let's just work on it together and you guys oftentimes make great suggestions so very chatty chat is what I like to see there like I'm keep looking at it and I I almost wish I could have it convert to voice right away so I could hear it in my head as you guys type which would be that's this next year right where we have the automatic translations so um so that's what we're going to kind of do today uh we do have by the way some um some updates as well I have found a node that will allow me to use a local large language model inside a comfy so instead of using the open AI version which is what I posted a couple months ago and I still use it from time to time or basically make a rough description like what's the most interesting thing I've seen at the suban abandon hospital and have cat GPT come up with some really creative answer and then give us the prompt and and render the image which is kind of fun it's like a treasure hunt in a lot of ways and I found a couple H ones over the time that I've done this but I've always wanted to kind of do it with a local language model I know you guys have said the same thing like I don't want to pay open AI or I don't have an ABI key how can I do this um this is a way to do that or get around that so uh that's kind of the the goal here so let's let's get going on it and we'll um I'll show you kind of what I'm going to do and obviously keep the questions flowing as we're doing this cuz it's it's as my it's as much my show as it is yours right um so let's switch over here uh so again thanks to gigabyte again for setting of the laptop it has been fantastic I have to tell you my favorite thing about this thing which is going to sound kind of weird is I don't know what this trackpad is made of but it's like freaking amazing like it it feels it's such a joy to use and it's such a dumb little thing to mention but uh also the keyboard on this thing is great the MSI keyboard they have a transparent key so they all glow this one just uses transparent letters but I had trouble kind of seeing the keyboard when the LED wasn't on so kind of an interesting usability thing anyway I won't dwell on that too much but I do want to mention them because they they are obviously going to be a star of the show here for a while um all right so I have two inference uh setups here we have one that is this is the laptop this is the one without the share button over here and then we have this one this is my local machine here that is using um obviously my PC now my PC is actually in the other room the only thing I do is feed through the the wall just my monitor cable and my keyboard and my mouse and what have you that way it keeps the sound floor kind of low in here for you know obviously making audio and video the thing about this I have to warn you though is this laptop like it's running real it's running really nice now and have it what's called normal mode but it has fans on it that can move this desk away from the wall uh so depending on what's going on this might get kind of loud now I really like the fact that I can turn the fans down but that will suffer performance so uh we'll have to kind of decide if it becomes annoying you need to say something you need to say hey this is uh this is too loud let me know and I'll turn the fans down and we'll just it'll take you know quarter second longer to render the image U but at least that way I'm not annoying you with the audio because the first thing I do is turn the video office because I hate the audio right I'd love to see the setting in uh focal camera focal angle and angle okay well like we'll show all the stuff here so let's start out with this so this is the image that I that I took this is an event called Texas school that takes place in um Addison Texas year and it's like 20 some odd um teachers and then each uh each teacher has about about 20 28 students or so and you have the same teacher every day for 8 to 10 hours a day and I'm one of those teachers and you usually teach photography Advanced composition uh lighting uh actually kind of even the Fine Art of Photography because a lot of photographers focus on the technical aspects but we don't get training in like why we do what we do what things you can use to emotionally move someone with an image and so on so this was obviously a lighting Tech demo and I'm showing this um this image was just a test image and I kind of passed it by but then I started thinking you know this would be an interesting one for dress design because it's a very simple dress design also shows some of the weaknesses with AI and our rendering capability so I thought that was a kind of an interesting thing for today at least I think it will show that again this is something I haven't done yet with this one so we'll see how it turns out but U we shall see uh looking at the um real quick looking at the chat here how do you do photo bashing and comfy uh UI uh multi- segment masking control net automatic yeah we're going to do some of that today actually so that's a that's a good time to ask that question AK creative agency and squirrel um it's all good Scott I think your vocal processing is uh cutting a lot of the background noise good awesome uh to use an efficient node for the a a IO prompt I do not um we're going to we're I I try and stay away from the efficiency prompts except for uh if you're referring to those nodes anyway for the um for the actual graph like I only use them for the XY plotting for the most part uh do you think the old AI models are going to be banned after application of the EU AI act I have no idea that's that is so beyond my pay grade I don't even think about it we have no effect on it so whatever it is we'll just deal with it you know there's like so many things in life you can sit and worry about the price of tea in China which affects a lot of your life you know how the the amount things are made how much they're made for importing cost all that kind of stuff but yet there's nothing you can do about it so don't freak out about it just whatever the decision is we'll just deal with it right hopefully they're being well educated which I'm not sure we'll see and then uh hopefully they'll make the right decision the thing about this that I go off in a political tiate I'm not going to talk about so much about the text to image properties but but AI in general the amount of really great um chemistry that has been done already we got 330,000 new Crystal formations have been discovered by AI That's like 800 man years of work to figure out all these formations those are used for making better memory making better chip designs making computers faster quantum computers all that kind of stuff just got a giant boost forward because of AI so to kind of come up with the hey we're going to ban it here doesn't mean that every country in the world's going to ban it so if the EU decides to ban AI in some way shape or form it doesn't mean that Russia won't or the United States won't or you know any other country so we could actually lose placement you know we could fall behind the rest of the world because we're trying to regulate it so much so we'll see how that turns out I'm I'm hopeful that there'll be education on that side but I won't go off in that too much anyway so I've done here is I've just exported this image so again it's it's not an overly complicated picture it's very simple um but I just kind of wanted to show you what it was so you knew what I was doing all right so on this laptop over here we have uh I haven't really done much of anything uh with it so again this one is going to be the local machine and this is the laptop uh the laptop does not have the share button so one of the cool things that I have uh in in my other one is is I can load up if we if we load the default for example we just get this thing which is fine for most of the time but if we really want to benefit from having the two prompt the positive prompts we need to use the other the other workflow so this is a workflow I've given you guys before it's in the community area here on YouTube uh and once again thank you to all the people who support the channel those links in there for those for these graphs are uh available in that community area and again for what five bucks a month you can download all of them and then leave right all those plus all the live streams you've done in the past this live stream will stay up for a couple days or a couple hours depending on how things go and then you can go ahead and and access all the past live streams we've done down there so again thank you to guys I think we're one member shy of 300 so if you join during the stream by the way your name will pop up on the screen and it'll be pretty awesome so somebody should do that make themselves famous that'd be pretty neat anyway what I want to do is I like this graph quite a bit it uses a couple of neat things first of all it uses this pipe here um so this image pipe is nice because it can just take this one pipe and Lead it away and continue working without having to drag all these other um conditionings and so on around so I like this case sampler from the impact pack again you can see where these are from uh I have a few quite a few custom nodes installed we're going to try and avoid using them uh as much as possible use the core of comfy but but there are some amazing nodes out there and I really like this this Dr Lieutenant data his stuff has been fantastic including the manager so I really want to support the guy his YouTube channel is uh no sound and it's pretty hard to follow but I'm trying to like follow up with his his content and add videos covering it uh so you guys can understand what he's doing because I think he's making great Headway in a lot of areas U but uh we're also using a global seed so this will replace the seed in every single one of these controls uh now I have it set to be fixed because as we're experimenting here we do not want to to have a moving Target right we want to set the Target and so we can kind of see we're doing with the nodes and not so much hope hope that oh my oh look at that we solved the problem with hands well no it just happened you got a random seed that look good this way we have a much better control on things we're only moving now eventually we'll move this to increment or something else you can also do uh on here there's there's some other ones too like you can increment increment for each node or randomize for each node independ ply I'm just going to leave them all set to nine and be done with it I like this workflow so what I did is I saved it off right I just hit save I just called it workflow that's fine now I went into this company this is on the laptop I can go and I can grab this and drag it down into here and now I have this now I can do save save as a workflow and we're going to call this sdxl base now whenever I bring open uh a new compute and I want to start uh if I don't don't have the default change which by the way you can change now um the python uh python Goos um group of nodes has the ability to replace the default um I don't know did I even have that installed on here I do so down here you can see what the default workflow would be and we can change it now to sdxl base and hit close all these additional ones with these strange icons here have been added because of that python Goos or python go he works for stability with me now too I I haven't really talked to the guy at all but um I talked to comfy quite often but uh I haven't really met this dude um but all these all these extra ones with the little snake are from that that uh add-on and this is one I really like because now when we go and we just hit load default and it says okay we're going to get this workflow every time because this is where I want to start this one is the The Point of Departure for everything we're going to be working on or everything that I work on um all right let me catch up on the stats here a little bit um been learning comfy oh awesome thank you thank you like like my live streams great uh by the way all the all the very Niche your content uh and the way you explain things is top natur well thank you very much I appreciate that I do have uh I do have one more video coming on the the the laser cutter in the Katon board if you haven't been following that it actually finished it and I gave it to my daughter for Christmas so it's all gone now I'm going to start a new one uh but I think it's be a Monopoly board but I think I might make one more Katon board too I want to do a steampunk one but I used AI to help me with ideation for those pieces and then I would create the pieces in laser are cutting software called light burn and then we take and burn them so it's actually I like the practical application and this is where I think the jobs are going to be available for people how you take what we're learning from this and applying it to practical applications manufacturing designs areas of ideation that's where this is going to win the whole imagery part of it is fantastic but again it's very easy to do and differentiation Market differentiation is important in every every job so let's find a way to make this work so that's what we're going to do today uh see by the way so is it possible to use advanced stuff with vanilla comy yeah all the time I'm trying to keep it clean with other nodes you know the thing about comy is the core the core system is very good right but comy only has so much time in the day and the idea was to make it so it could be expanded so don't think of it like you're dirtying the product by adding custom nodes it's intended to be that way right it's a backend first and a frontend second all right so let's uh let's play with this thing today so again we're using the laptop here I'm just using an internal IP address and when I launched comfy I launched it with a different um I covered it last live stream I can break it out if you need me to but I basically remove the portable thing and put the word listen in there and uh that's it now this laptop is going to be doing all the all the work and I'll go to the screen from that from time to time so you can see it all right so what I want to do is I want to load in uh we're going do image load and we're going to load in not that we're going to load in the picture of that girl uh so I'm just going to grabb it and drink all right I did I did create some fun little um things by the way I have a friend I was doing uh I was trying to make like cute little dowels of things I made this of her um and I this whole this is just all prompt mechanics by the way I'm not really using um any sort of a a Laura or anything like that I try and avoid using luras for most part uh because the the base model is very flexible it you can make you can prompt usually what you want unless you're looking for something really specific in which case that's when the lore should be used uh looking at the workflows that people are posting for the uh open art. competition which we're I think the finals are being discussed tomorrow and then the winner will be announced or winners I should say uh the the idea that you load every Lura you can have just to load it is a bad strategy by the way and I see a lot of people doing that I see like some of these workflows have 30 luras being loaded up like what are you doing exactly you just just replacing the blank model or the the base model all right so what I want to do ultimately here I want to see if we can be successful is we want to replace this dress with other dress designs right so that's going to be our goal so we're going to start from this so the first thing we're going to need to do is figure out how we can do this so let's let's fail first and work our way forward to kind of show where the strengths and weaknesses of things are hey woohoo Charles you are probably number 300 there I think you're three number 300 on the membership chain thank you very much sir all right so uh one of the things we need to do is we need to work forward so I'm going to take and kind of work my way through the problem solving of this and uh we we'll probably fail Ian we're going to fail initially I'm sure because what we're going to do is probably the wrong direction initially but let's look at the strengths and weaknesses of each node as we move forward got it yeah you are 300 thank you sir all right so the first thing I would probably do is let's look at uh what would we do if we just wanted to Simply take this and and uh we're just going to use this as a an in denoising process well that's probally not going to work very well because it's just going to denoise it we want her to stay the same by the way I want the same pose I want the same girl effectively for every one of these now later on I don't mind it so we have to change her hair color or things along that lines but for the most part I want her pose to be the same think of it like you're trying to sell it to a dress designer and say I've come up with 30 or 300 or 30,000 dress designs for you based on this they're not going to sit there and go wow these poses are really interesting what's this no no I want you to look at the dresses not the poses so again eliminate variables static pose for each one how do we do it and we're not going to use open pose in case you're thinking that's the control n I'm reaching for that's not my plan now and we may have to go that way but that's not my initial reaction to thing by the way if you like this stream and you're enjoying this please click the like button it's really interesting you you kind of forget the fact that's kind of like the way to clap to these things so if you enjoy this please click the like button since 100 people watching and I have 25 likes so 75 people I'm either putting you to sleep already or you forgot about the like button so I appreciate that all right so how do we do that well we want to replace this latent with this image right so I'm going to duplicate this image uh just because we know we're going to need it a couple different times here so how would we do that obviously we're going to re-encode it back into a a regular image right so or a regular latent so here we go and now we can replace place this up here and we're good to go now from here we can go ahead and decode it back into a regular image like just like you would expect very very simple now let's talk about this for a second here what are the things I do not know about this and and don't come up with a solution that works for one image like hey this works great just for this one image uh what what's wrong with my workflow well I don't I can't trust that the image you're handing me is actually going to work right I don't know that it's going to work and sdxl has is some sizes it works better than others and I'm kind of even going to ignore that part of this for this at this time but how big is this image I don't know it could be 4 gig it could be 25,000 my camera shoots 40,000 pixels I think on the lung Edge that's too big right we can't do that so let's change this and there's a really great again Basse node you just type in edge here you can see oops or side sorry side there we go image scale to side and I like this node quite a bit pop this in here and then we can say that I want the longest side to be 1024 there we go this helps kind of protect us from someone handing us an image and we're trying to demo hey here's a where you can do great dress designs well here's an image I'd like you to use okay and then everything is really slow and they're like wow we don't really want to buy your service or product this is dumb uh well the reason is because this image was too damn big so let's fix it first and you can see where this comes from this is from the deru not no set again there there are some the impact set the W node Suite this deru set and there's uh MTB I think it's called that's another really good one that has so many core components that other people are using that I think that they're they're kind of mandatory like you have to have them because other node designers or developers are taking bits and pieces they're crediting them and and by the way I sure hope that all of these people eventually learn how to turn on the GitHub uh donation page because I give to the people who who help me do my job better right I mean just like the people who are paying to be members of this channel I appreciate you and I want to give that appreciation back to these people here you know I I paid I paid uham stability when I would use dream Studio before I was you employed by them I paid for that product I pay I still pay for Mid Journey although I haven't used it in in a a while um I do like version six a bit but um it's it's not uh I think the the direction they're going is interesting it's it's uh more photographic right which I know a lot of people are really into I'm not into that I I don't need photo realistic part in my life because I have a camera and I get hired to use that camera and I I love the artistic aspect of it though that's the parts that are refreshing to me the parts I really like all right let me catch up on comments here a little bit um uh you found one node all the time that is image to image is save from WS node sweet nice to be able to control the save file yeah so they have a there's a a much nicer save in there uh the W node Suite has a lot of great image manipulation tools post-p production tools like contrast and black and white adjustments and so on a lot of great stuff in there um uh what's the node you can only control one side of the size the image but not both the height and width yeah that's this one here that's this image scale to side so this the length of the longest side uh that's really all I'm doing here and I'm not cropping it after that I'm just kind of scaling it or or whatever here I actually used the nearest exact so so that's what we're doing there okay so will this work okay and let's let's look at the prompt here this prompt is what we had in from before so let's type in uh um let's do um ballerina well I need a prompt from you guys too I'd like a prompt for dress design say female fale with red hair Coral this is the color coral by the way that dress dress with stripes blue stripes let's do yeah blue stripes we do blue stripes because it's really obvious if we get it wrong right because blue stripes on the coral dress are going to look pretty hideous uh so there we go and uh yeah let's see what happens oh we have the longest edge here is 1152 so let's actually change this 11 match um because I am using this sdxl resolution again this comfy math node Suite f fantastic node site I use it all the time are you involved with the hug course in stability in January not yet there's so many parts of stability that I'm involved with so I did quality assurance basically for all the modalities but some more than others like the my passions are and and the things that I do um different things along that lines but you know I like for example I'm really big into the boml part too I love the chemistry and physics part of what we're doing there that is going to change the world this is also going to change the world world right but that's going to like make us live longer to enjoy this even more right uh it's a great combination of things all right so if we if we do this the the thing we got to pay attention to here is how much we're denoising that original image right so if we denoise it 100% that means we're probably not going to end up with much of anything let's just try it and see what happens again we all the reference is being handled by this laptop over here uh so we can switch to that at any time if you want to see it oh I'm using a little screen here I got to switch to the the big one first for my face to this so this is the laptop's thing already [Music] done and Valu not in list looks like I have to go and oh I did not update I did not update the S I got too many switches here I'm trying to figure out what the hell I'm doing I did not update where this model is so I'm just going to use let's use Juggernaut today fun model and I think version version s he just announced and Redd it is about to come out so whatever he's making there over at this laptop to see running and now the fans are going to start to kick in a little bit so you let me know if it's a problem well it's because it had the wrong path right that was not the laptop's fault it was my fault so hey look at that it is a redheaded girl just like we asked for and it has absolutely nothing to do with with what we wanted um so let's change by the way the where this um feed is going here FL Gallery where is we want this on the left I'm just going to move it over there so we can kind of see what it's doing oh look at that learning core just became a new member wo thank you all right so we're going to make this smaller and stick over here all right so this is not help right uh it is not the uh it's not the girl it's not the dress and every time we click this button if we have a different seed uh we're going to end up with a completely different girl that's not helping our process at all right I don't know what happened here but looks like it's the same girl but that's a weird picture all right so let's fix that how would we fix that well the noising process here is probably the first thing people would reach for see well what if I just 70% D noise this okay better but again the pose is wrong it's not helping us out real real well so we're having issues here and obviously you're you think you know where I'm going here and this is where I'm going to go is I'm going to go with a control net we're going to start with a control net and that should solve hopefully a lot of our problem but I'm going to do it in a way I don't think you expect them to do so let's add that in so I'm going to first of all I'm going to go back to probably just the regular empty latent here um just for the time being so if I if I click prompt thep again here you can see that has absolutely nothing to do um with this picture they're [Music] unrelated get rid of all my copies this all right so let's start with the control net so uh and we'll see if I have all the control net on this machine by the way if that fan is too loud you need to let me know right it's not really loud right now I can put it in meeting mode and it'll be a lot quieter um but we'll see all right so control net apply actually I'm going to use the advanced I like the advanced one so the difference between the Advanced one and the regular one is we can we can determine when we start and when we end as well as the overall strength of it that is very cool and you don't think about that if you hold say the um the cany which is the outside edges of things and you're holding it at 100% the entire duration you're not going to get a lot of wiggle room in the model but maybe you want a little bit of wiggle room you want to let the model be creative or maybe you have some luras and you want those luras to be able to chance to have a CH chance to shine if you don't have any wiggle room in the start and stop you're not going to have that wiggle room right so we want to have this this Advanced one and this is the one I use pretty much all the time I don't really use the other one at all now again I'm not sure if I have these things in here but let's we'll see how it goes all right so uh conditioning wise I need to pull the conditioning up for the positive and you'll notice that the advanced control net also uses the negative usually a control net doesn't need the negative but the advanced one does use it so we're just going to go ahead and let it happen go our positives and our negatives so again we're just just putting it right here again this is the reason I love comfy is we now know where the control net sits it's not a pixel level thing it's not a model manipulation thing it sits in the conditioning pieces of this and again comfy says you know if you if you talk to the guy or read his blog that he wrote this to understand how how stable diffusion works right so it's not that we're becoming great prompt engineers and you know automatic 1111 is a great product and I have it still installed Al I haven't used it in a while and it does some things better because it's easier to use in some cases like training for example I'd rather do training actually with a command line than I would using a UI just because it's resources but you're really kind of open you know uh you're kind of shut down and you're not open to how it actually functions and I think you're just checking the use aora box and you pick it you pick the weight that's great but you now know where this goes which I think is more powerful right uh as far as an image goes let's just use this one but we can't just use this one in its current state right what's the control net we're going to apply let's go load in some control Nets here and I put in just two today I put in a cany and I put in the depth right so let's start with canny and start with that one now in order to use canny I need to actually hand it the the pre-processed version of this so if I just type in canny you can see that there's a few there's one that comes with Inspire and this is for segs this is for segmentation of out of different things we're not going to be doing that at this point uh there's one here and one here and there's there's three look at that so there's one in the core uh what are these other ones here I'm not even sure this one comes with the w node suite and it has uh looks like uh another thresholding capability and a different default which is interesting and there's another one here too what's this one this is from uh comfy UI control net um this is the alternative control net I think the auxiliary control Nets this one same idea but this one has a resolution on it which is he hey that's really interesting let's try that one I haven't tried that one before what's our resolution for 1152 let's use that that should give us a much more accurate outline at least that's what I would imagine this does so in order to find out let's just go ahead and do a preview off of it and do a q prompt real quick and then stop the prompt see is what we get so that's that's very nice it has all the finger joints um outline of the dress for the most part uh again we can affect the high and low resolutions here to see if we can get a better shot of the dress here so we want to keep these up or down we're just going to play with them okay we want to go down until we get what we think is a good outline of the dress I want to see her arms for example that she has two arms and if I go with too much detail by the way it's going to start the image is going to start to suffer because again that wiggle room is going to go away but I think this is a this is really kind of an interesting start so let's just start here and see where it goes and I'm going to go with this pre-processor here so this came this comes with the auxiliary control net models I believe uh so again if we go into manager all custom nodes this again this is the laptop so there are not a lot of custom nodes in here I haven't gone nuts on it yet like I install a lot of stuff just to try it uh Sebastian on his channel does the same thing he installs a lot of neat nodes and tries them plays with them and there's a lot of great discovery to be made in there um I don't have so much the time to do a lot of that um but every so often I'll find one he's he's been a really good resource for stuff by the way so there's so many good channels out there um Mato the guy who wrote the uh IP adapter his his Channel's just starting up he's got I think five or 6,000 members already too so that's growing great um great guy too work with him uh we were paired up to judge the open art. a competition which again I think tomorrow's the last day of judging and then this will go to the awards but there's this control net auxiliary Pro processors this is where we're getting that node from that's that so this is what we're going to use we're going to use this we're just going to plug that into image here because we need to have it pre-processed in this canny this is the canny method there's another one called H which is kind of a soft outline of things um it's a little more um gestural I would say but not as accurate okay so so let's just try this and see what happens right so if we just cue this prompt we still have the doo set low so let's turn the doo up to 100% again I'm just going to cancel it [Music] here the laptop is starting to kind of build frisky over there and I didn't change any of these settings so let's do uh DPM 2m Caris here that works really well for sdxl and 20 steps is fine CFG of8 is fine and we end up with that all right so I'm going to move things around a little bit here and just so we can kind of see I'm going to borrow this image again and put it over here just because it's our Baseline for everything and how we're going to judge success okay so remember the the mockup here the idea is we are going to a company that wants us to they want to hire us to do the AI to help them generate dress designs so they want as many dress designs for their upcoming line as possible uh using the same form factor but they're looking at inventing different text textures or textiles right all right so how did we do how did we do so far so uh interesting we have something going on down here and this is uh this is an interesting part this this is one of those things that I don't know how to get past right it's uh it's going to be a a problem for a while in my mind unless somebody's got great ideas and you have to try them don't just like throw them in the chat cuz trust me I've thought this through but I've running this a numerous times is if it can't see the feet it assumes the feet are missing and it needs to add the feet now we're going to try another technique to avoid that but um I think we're going to run run into it pretty much all the time so um in this case it shows you if we're going to do this type of situation we try and get the feet into the shot so that the the models all work out and we don't end up with any weird surprises anyway so let's try another in fact I'm going to let's let's put this on an increment in fact um I was thinking that I would turn this to Turbo and see if we can get it to work in Turbo but I'm not sure um if that's going to function or not but let's try it just because it be kind of neat we can these out very quickly now the turbo model is going to be a 512x 512 model so we may end up asking yeah we may end up causing ourselves some let's not even try it then I was just thinking that would be kind of cool but I also have the dream shaper mod uh dream shaper model which I just downloaded this morning I've never played with it before um but uh there's a few other ones this uh realistic Vision um I again haven't I don't think I've played with them I think I played with that one early on and it didn't have a vae with it so it was kind of weird but again you guys you guys probably deal with this stuff more than I do um I I use the base models there somebody posted on Reddit they're like hey we should have a competition and say who can make the best image with the base sdxl 1.5 I'm like I'll play that game all day long like I have so many of my greatest I think my greatest pictures are sdxl 1.5 vanilla with no lauras or anything I'm really liking that all right so that's okay she she uh she needs some sun now or some vitamin D supplements I think but uh yeah otherwise it looks like we're we're doing okay again is trying to add the feet uh but this is looking pretty good now we talked earlier about this um I'm just going to I'm going to move the controls around here so we can kind of see what we're doing let's talk about the strength of this so let's look at ending it early so let's let's let's end this at say 75% of the way through what will it do so it should create use the cany first hold the line and then let it go go right well now she flipped around that's interesting so do we need to put that in our prompt [Music] maybe facing away from camera don't know if that's going to make a differ again so you see it get through about 75% of this and then it just kind of does its own thing like for here this is really interesting it cut hold through the dress here because it could so that's not a huge difference so let's keep the UN percentage the same let's change oops let's change the beginning so this means that the model can do anything it wants to for the first 25% of this right anything we'll see what happens yeah you think the negative of no visible feet now it won't do anything guarantee you uh the reason is because it doesn't think about that um fact have iron no feet or just feet Feet Shoes legs any of those things right [Music] um it doesn't work that way but so you can see here it went off on a wild tangent for the first 25% and then it tried to bring in the control net and hold this which which is pretty weird but um there is a lot to be said for letting it do this now not so much with cany but let me show you in a second with depth because it's really interesting so again you can see it started out really weird and then it kind of came back around hey that time it worked look at that isn't that funny if that prompt would to be the thing that actually works again we can't say that oh my God we've solved the problem use this and your negative prompt forever it might just be the seed right yeah it might just be the seed maybe he did it maybe it's her entire damn lower body that I put in there but when we hit prompt if you watch this watch that it's going to draw a completely different image and then it's going to use those pieces of image to try and glue together yeah see it didn't we still get shoes and whatever this is but it's that freedom of that first 10% or whatever that allows it to kind of go off on a weird tangent and then try and bring it back around why that's interesting is not so much with this control net but with another one so let's uh let's look at that one going to break this apart back over here we can we can actually group these things add a group here we can kind of like doing this my way of doing things now you can do you can right click now and then go to where is it used it in a bunch in the past but it's in order here well there was one in here that allowed the group to snap to all of these but uh I can't find it right and that might be an option again that's added by another or is it in here here it is fit group to nodes there we go that's a little tight okay but there's this one all right so I like this one this one's fine but let's do a different one Let's uh let's do a different control net so I'm just going to duplicate this control net here but we're going to use a different one we're going to use control net lower we're going to use the depth one I think the depth one is probably the coolest control net one of the coolest control Nets from an actual functionality standpoint of any of the other ones like I just think it's got so much more going for it capability wise all right so uh let's do depth now we've got a whole bunch of let's type why don't we dep we've got minus Lis Zoe uh midus is probably the most accurate but it's the slowest uh in my in my uh my experience so let's just try it just do an image here and then we're going to again if we're using a control net that is going to be doing depth we require a depth estimation to be done first and this is from the W node Suite I found this one to be very slow but very good right so get what you get what you pay for more or less I'm going to disconnect these things and let's just run our positive am I doing here that's a two from here need from the positive from the negative again because it is an advanced one do the [Music] positive don't don't mix those up by the way i' I've done the the negative to the positive before and you get some sort of totally bizarre thing I'm just going to disconnect these completely so you can see that they're not in play right so this is the this is what we're dealing with and let's again let's do a a preview so you can see what the depth estimation looks like and I'm going to leave these again set to one and uh we're going to start at one start at zero and end at 100 just like we would expect it to and again it's taking its time and it looks like it died did all right I don't think I have the minus depth map on here so I didn't think that through let's do a depth another one again I haven't tried this yet so let's do the Lis [Music] one at preview all right that one's pulling the that one's pulling the model down at least the other one did not me check them on your cont comments here um are there any nodes to save and play with Vector images not that I've not that I found I haven't really gone looking for them but I would like to at some point because again my laser cutter uses an SVG file so that's the best bet by the way we have 150 people watching but only 64 likes so if you like this and you like what I'm doing here please click the like button just show your appreciation that way okay um talking about Samplers do you activate uh the restart sampler I have I have used the restart restart the actual paper on restart sampling is actually very interesting so there's uh there's some stuff to be again I'm trying not to move too many things today so it doesn't get too complicated but um is there a new version of sdxl uh not not publicly available there's 1.6 that's available uh on our API but it is not a released model at this point I do not know dates on any of that stuff so I can't tell you um thanks for videos I learned so much well thank you for watching appreciate it it's a cool dress I think so too uh full length no visible feet yeah you can see it worked um I'm not really sure if that's uh if if we win or not we'll find out but this depth map actually helps with that too usually um yep there's our depth preview this is sdxl we're using sdxl all the time we're not using 1.5 for anything at this point and a lot of people like well 1.5 is is so much better 1.5 has been out a lot longer right there's a lot more support for it there's a lot more models or a lot more luras or a lot more stuff so yes of course it's going to be better for most things but um not for uh not for anything practical moving forward we want to stay on the bleeding edge of stuff if you're using 1.5 you're already almost a year behind right so why would you do that uh just stay with current stay with current will there be a 1.5 turbo I don't think so I mean I we we've called 1.5 models deprecated in our mind so I don't think we'll ever see anything from 1.5 that'll be added on but again it's not I'm not the strategist for the company but in my mind it's it's deprecated so I'm moving past it yes it's cool model and yes I still love what it creates again some of my favorite images are 1.5 but that's that's just what it is okay here we go um this is looking pretty good we are now producing dresses which is nice um we can uh augment our prompt a little bit so let's go over here and look at this we have red hair a long a look at that we had word along in there a a let's just get rid of the word aong so I didn't know what it was doing anyway let's let's replace this with uh some crop so red blue [Music] green stripes right and we're going to put these inside of braces separated by pipes and that will mean it will pick a random one each time right uh and that's uh that's interesting we'll just kind of do that I do like some of the Wild Card nodes and I have installed Wild Card nodes on my main machine but again where you're using the laptop to do all the work here and you see it's pretty fast where we got uh sorry I'm talking I'm turn my face with from the mic here uh we're getting these images in 10 10 to 12 seconds uh so that's pretty decent that's that's faster than my 3090 on this machine and this is a 4080 this laptop is a 4080 in it so it's working very quickly let's do this let's cue this prompt again it has to go all the way back here to start each time now because it has to go get our information on the random so depth I think depth is probably one of the most powerful if not the most powerful control net that's out there now you see here's the fault withd depth is she's on a blank background with a plain floor and everything is generated is going to look like this right you see the depth of the floor you can see the problem we're going to be having here right but it's doing so many other positives what can we do to kind of mitigate this well again some of this is let's let's bring this control down here some of this has to do with our starts and our stops and things net lines now a lot of people would reach right for strength and again if you're using regular control net that's the only thing you have is a strength let's say okay let's just apply this at say 75% uh let's do it again so 75% depth you if we watch let's start over again if you watch this this as it generates it um you're going to see from from point one it's already has the depth set right but it's being a little more Whimsical with how it's a applying it think of it like it's 75% transparent so it's not quite holding the floor but it's holding her fingers well kind of kind of now we got some SC crunchiness going on in here and her figure is different uh hair's okay the dress is completely different form so I'm kind of screwing up the picture by playing with this strength so if I leave the strength at 100 but I changed the start and this is I think where it's really interesting let's say we start Chang this to 30% if we watch the first third of the images being generated by the preview over here you're going to see how unique they look until the model snaps it back in this is the kind of the the way I WP my brain around it it's going to render a woman wearing a dress with red hair right it's going to render her and she's going to be in the picture guaranteed and it starts and it starts and all a sudden says oh crap I need to move what I'm making and shove it into place over the top of where that girl is actually standing so if we do that or we think with that mindset this actually seems to work that way so here we go and you see all of a sudden yo it'll try and yanker back into that depth position should what are you doing start percentage all right that didn't work why did that not work that is not what I expected to see so star percentage should have been right H that is yeah so that worked that's not what I expect when I expect when I change this start percent that's what I expect to see if I change it to 50 I should expect to see it change about halfway through it should snap her into the other place isn't doing it's interesting I have used this thing so many times to do oh so many random things so let's let's change it the other direction we'll start from the beginning it we'll go about halfway through so it should immediately start start and about halfway through it's going to go off in another [Music] tangent and we're not leaving in a lot of wiggle room here because these most of the remaining parts that might be the reason this is failing most of the remaining parts of this are already determined by the time it gets to half right so let's do that again but let's change the start percentage to something like 10% see how that works so the first 10% should be pretty weird and then it should snap her yeah there we go and you can see what it did let's start that again it's going to start drawing a picture and then about halfway or 10% through it has to snap and move what it's doing or try and move what it's doing which means we end up with the rest of the dress or some other part that was hanging around from the first time so there we go that was the problem I was starting I was being too aggressive with my start so 10 is pretty good let's move it up a little bit uh by the way you can scrub on these lines if you didn't know that uh so if you just want to change something instead of clicking this all the time you can actually just scrub let's do 15% see what it's going to do you kind of see what it did it started and then it kind of said okay I have to use this my my Point of Departure now what can I do with it uh and we end up with something that's just kind of interesting but what this is allowing to do is that the depth map is no longer being held responsible for that floor right but it is but not right away so we end up with a lot more freedom than if we hold this thing at zero and 100 this is this is interesting let's change this to 20% and about I think as far as we're going to be able to take it 20% in yep has to try and fit her in there and it's going to fail because it's already too committed so that was let's do that again that was kind of an interesting showing of what's going on here so it starts and it tries to bring in the other one and you see it fail So 20% is too much let this back down to that laptop being too loud I'm so happy with this thing by the way like I bought that MSI laptop and I was like really happy to finally have a new laptop and it was like was I it was never quite happy wow we've made some sort of horrific thing okay well I'm saving that one we have a uh we have a at work we have a cursed art share and uh yeah this is going in there so again 15 looks like it's a bit aggressive maybe 12% so how much wiggle room can I give it before it has to change all right 12% looks like it did okay kind of well kind of and I'm I'm showing you this because I think this is a really great way to to understand what the depth map is doing as far as holding or trying to hold this thing in place so I think 10% is probably 10 here .1 10% that should be again enough to let it start on some wild adventure give us an interesting background or a scene and then go from there so now if our prompt says something like like [Music] um on the water's on the water let's move it up here though red head red hair on the water and whatever color dress with Let's do let's take all these Stripes here put it go so we're going to get different colored stripes on a different colored dress again I have wild cards I have huge wild card files that are really fun to play with but um I don't want to kind of don't want to clutter the live stream with too much stuff kind of keep it a little bit clean by the way I like this one like it gave her a tattoo I was just noticing that I'm like got a tattoo right oh yeah before and after I want to keep those next okay how are we doing on the the the concept of the pose not changing it's pretty good right the depth map is holding it pretty good but we are getting some movement in the hair that's different each time which I think is really nice but we can control that if we need to but again I've just shown that my depth map is not really the place that I want to do that I think I'm really happy with this setting for this situation so I'm going to leave it alone in fact again I'm going to kind of Tuck this stuff together and let's make a group for it and then oops on the group whole group management part of this needs a little bit of love but again this is the tiny nodes project not the comfy project that that's responsible for the so we'll just say this is our depth control net here now it's a group I like to group things because then they're easy to remove in fact uh if I was doing this in a competition like if there were a competition right now I would actually make nodes like this and park them right on the edge this is how I prefer to create these node groups because then as I'm using these different positives and negatives for [Music] example this positive to the positive here and this NE here I like them on the edge like this because then as I move this around it takes them with me but I can save this group off and use it again later uh in other workflow so I I tend to do things like this um so that I can bring them back those are templates by the way so if you highlight a group of nodes like this you can right click don't right click on the nodes you have to do it off to the side you can say node templates and then manage or save selected as template and then we can say this as our um net yes by the way when you name things a good idea is to name the family first and then what makes it unique otherwise you end up with things in non-alphabetical order so all my control Nets will stay together now so control net depth control net cany control net whatever that's just a good rule in life so now when I'm doing this I can do node templates and control net depth and now added it right in here it does not bring the group in with it I wish it did but um currently that is not something it does oh well so uh I think this is working great but can we add back in this capability here CU I liked the the ability for the fingers to actually have the individual joints the way they're supposed to be things along that lines and sure enough we can we can run this um the out of this one right we can run it into the in of this one so you just chain them together it's really not anything special right the positives here we go to the positive here all the way down right now we have two chain together things we're doing a depth map first and then the cany second again the order is not important as far as I know the order is not important I could be I could be wrong the order might be important but to my knowledge order of this operation is not important um some there are several that are like in catenation of prom s or and conditioning do matter but in this case it doesn't matter let's cue this prompt again and see what happens again this and you see 20% was allowed to kind of do whatever wanted to do and then it took over and you see now we're on some water and we have lovely weeds in the background and we get this horrific dress that we have we're going to try and sell this dress and now the fingers look much better they look like the original girl remember this is a photo I took in Texas right it's not an AI generated thing and I think a lot of us get stuck in the mindset of I'm working in AI therefore everything I'm doing must be AI based not the case at all this is a photo and I'm working with a real world situation of I want again hypothetically if you're just joining late hypothetically I want to design a dress based on this current dress but I want different patterns different ideas and a firm has hired me to give them this these ideas they don't know you know they're stuck in a creative block or they want to come up with a new line and they have a set of of colors perhaps they want to work with what do they want to do and I'm just offering different now let's make the prompt more interesting I'm going to drag The Prompt over here I know I'm kind of making a my own UI but that's the way that I use this system when I'm it during the day is I just drag my nodes to where I want them to be and I use them way so I just zoom in and this is where I'm going to be working so um I want uh so it's a dress um I want to say a butterfly but butterfly dress with do we'll change [Music] this and we we're doing some things with uh how to manage the you know the feet uh and we came with a negative prompt here um that I didn't think it would actually work but I did things like even her entire lower body I'm sure this is not correct but it was just funny that it worked and this is where bloodletting happened right in the Middle Ages well we bled him and he got better so if you remove blood from so get better right this is the exact same science that's being applied here do not take this and say this works every time it's just dumb luck that it worked but what part of it made it work we don't know we'd have to take it apart and kind of figure it out and it might just be that this part doesn't need it we just get rid of this um but yeah let's do that let's just get rid of this part cuz I don't want to I don't want to be that guy who's like oh look you invented a giant negative prompt from his y yeah so it's just as good without that snide comment in there about her entire lower body being weird dots all over the picture now that's interesting and we got to wonder why that happened right me put now the dress is uh a little weird right why is it weird what's going on well we what did we change we added the the the control net for the canny and again it's adding this let's just move these previews down so they kind of we're kind of all working together so you kind of see what it's doing right this is the family of items we're dealing with we have an outline of the dress we have a depth map of the dress or the scene I should say now [Music] just these are the two things that we're holding as true we're not letting this be true until 10% of the way into the image so we can go ahead and mess with the floor or otherwise we're going to end up with a blank room right it's just going to be a box just like it was shot on just a sweep of paper and uh these are the outlines that we're keeping now right now we're keeping the outlines 100% of the time and they are they're true like that's all we can do and we can go up here and we can play with this again like what's the strength of these dots how do we apply that and maybe we drop it down to 80% or maybe we start late or maybe we end late I think that's more interesting again maybe we stop forcing where the fingers are and everything maybe 80 we can probably even maybe do 60% of the way into this scene because by then all of the different areas of her body will have been established right but maybe we can get more hair movement or things like as you see we are holding the hair in position with these lines right so let's say 60% of the way through the image this is true and then afterwards go ahead and let it fly and that's kind of what this is doing so let's see how that works so you see we're still applying it we're still holding that true and then about here it's going to just kind of do what it wants well it's obviously pretty much already established at that point so maybe 60 is again too big a number let's borrow this control bring it down and I can rename this one too uh if I need to so it's called by the way uh once we're done with this adventure today U this workf will be available for all the channel members the sponsor level or better so it will be in the uh Google Drive that I have for you guys so thanks for everybody for supporting this I really appreciate it Scott is on his grind right now yeah I am is a cany uh yeah so this uh this work will be available for everybody supporting the channel so I really appreciate you guys so I try and make sure you get these workflows at the end uh it may not be the most colorful thing in the world you like well that wasn't really you know super in-depth and I could certainly recreate it but again it's what I have right here you go all right so let's say let's say about about halfway through it can stop ignoring it can start ignoring this cany let's try that again is doing say you get a few like I would appreciate an upscale and to clean it up more see the skin is not so great yep we can we can do that oh there we go but for out of the box that's pretty darn good right our fingers are good even though they're complicated they they look good we have basically the same woman although the red hair has obviously changed and I'm not going to pick that apart we could come up with whatever color that is but we're getting a really interesting scene uh the trouble with this in my mind is this depth map again is it's the King right it's doing so well but it's also representing the floor here which I may or may not love but what else can we do with this thing so let's let's talk about masking and somebody called it they're like hey you could do masking by the way I'm going to abandon this whole thing down here I was I was showing that earlier we're not going to do anything with that um I talked about that we were going to start out with a fail and and that was going to be de let's talk about how to m ask this off now when I when I do projects like this I don't want to be um I don't want it to be something that says this works for this one picture and this only one picture it's like oh every time you want to edit this all you need to do is right click and then go into the mask editor and and edit and then edit your mask using the the clunky tools that are in here right no one is going to tell you that that's a great solution for we need to do this in an automatic method right one of my favorite ones I didn't even think about that it may not be on this laptop we're going to find out again thep laptop the uh gigabyte 17 XO here is doing all the heavy lifting for the for the show today so again thanks to gigabyte for for throwing this at me um fantastic fantastic product again I put the I'll put the link to it down below um so you guys can see it once the show is over so we're not even using local machine at all today but if I needed to I'm keeping it open because if I need to save you know save the show because it's going off the rails I can save this workflow and drag it right into this and we can continue working on the local machine um that's one of the beautiful things about working with comfies you just drag and drop the images or anything and it'll work well so for this I want to see if we have ultral litics on here we do so ultr litic uh which is installed with the impact pack by default is a fantastic set of tools and there's it comes with three but there are many more I probably should have thought that through we're going to use the person and you see it set seg M person and there's bbox face and B box box hand you have to make sure you're using the right one here so segm person which means we have to use the segm connector if we use the bbox one with the segm model it's going to break which makes sense right uh it's telling you perfectly what it needs but there is no other information there so what do we want to do with this thing I drag out the seg m and just let it go and it's going to have some options here we're just going to choose the combined B2 right I'm going to drag the image that we're trying to use is here and then drag mask and I going to do is I'm going to I'm going to drag the mask to I want to preview the mask and I don't have oh yeah mask to image let's do this mask to image and then preview let's just run it real quick and see what it looks [Music] like should be good like the laptop here and see what it's doing I think it has to download the model cuz again I haven't done this before so um new laptop new new models new things to [Music] download yeah I don't know what it's doing what the fans are going like mad well while that's working let's talk a bit about everyone who's making this thing happen huge thanks to everybody who's supporting the show today we just passed 300 members so I really appreciate you guys doing guys and girls uh really helping to keep the show alive appreciate you so much um and again all the files for today and all the previous episodes and previous workflows are all available plus all the other stuff that I throw into the community area new links to things I just gave out a new um just gave out a new workbook to a a collab notebook to use the uh new stable diffusion video API and again even if you have no intention using that API it is a great way to learn the rest API and see how we're doing things but I will tell you that there are already clients who are looking for people to help them learn how to use that in their businesses so you now have a tool that can do it for you you don't require a lot of heavy lifting is you know a laptop or a big machine or anything because you're using our API so you sell your services to companies and yeah you're going to pay what pennies fractions of pennies for our video API but then you don't have to buy a you know $4,000 laptop to do it either all right so this finished here you can see that we got something out of it it's okay but what's it going to do for us so we can use this mask potentially uh to do some of the heavy lifting of trying to keep this thing together so if we we we kind of hinted at what we're going to want here is if I drag this mask I'm just going to drag it down over here and there is a vae incode for in paint now this has its own problems and we're going to see those in a second here but this is a really interesting idea right we're going to inpaint everything that's white so drag it up here to our latent because that's what we're going to use latent we need our vae here of course and we need the actual image the pixels of the image so we're I don't know why I put it down there that's where I wanted it but maybe I maybe I don't put it right you can see what it's doing now it's going to grow The Mask by six pixels and then try and imp paint now what's wrong with this picture well what part of this are we trying to imp paint right we're trying to imp paint the girl not this so if we if we run this prompt here we're going to end up with the opposite um it's going to what going to find out here yeah it's keeping the crappy background in the studio of the original image and then changing the girl completely but because we have the depth net the depth control net and the cany control net uh holding it in place we're getting something that's kind of passable but we're going to get all kinds of tearing right in through here because of the roughness of the Mask this is the one reason why the inpainting node or the latent of the the in paint is not it's not one of my favorite methods this control net method is cleaner but this is again another option so if we if we take this we need to invert this mask right so this mask is backwards so in here we can pull this down it doesn't appear there but if we need type the word invert we can see that invert mask is so we just drag the mask into this and now we have an inverted mask we can prove that by obviously throwing it and we need to take the inverted one down to where it's actually being used so now this mask will be opposite flipped and now we do it we'll end up with a better solution but we still may get that tearing in and around in and around the girl and that's because again the resolution of the Mask is not as good as the resolution of the image yeah you can see all the tearing that we have in here is just trying to deal with that uh and in the background so U it's holding her so that we're not changing her at all right but we're changing everything else around it are we piping the full size image in for segmentation uh we are not we are using a uh a node to limit the edge the edge of the of the image we were using it well maybe I didn't put it on this one I had it in there um so image scale to side I will tell you though that I know this image is not obnoxious like when I when I clipped it out I think I did it at 1024 um so I know that's not the issue get used to substance substance designer has a lot of really great tools for node manipulation and um the people who make the light nodes project i' learn a lot from them as far as usability but who knows if that's ever going to so in here and we can do what size do we want this we want it and did have there it is [Music] 11 now let just it's working but it's not fantastic right this mask tends to look kind of crappy but we'll see if we can get this to be a little better so try it again with that side thing but um I think we're going to put the same tearing this just seems to be part of the way imp painting works from uh this these edges these edges are just brutal plus we have a gap in here and if you look at the mask you can see that we indeed have issues with the mask mask is not accurate anymore it's failing in fact it's failing in a hard way this is not very good at all so we need to find a different way to do this um there's a combined things to man no there is another one in here have to use the SE [Music] to yeah this is going down the rabbit hole for where I don't want to take this but could [Music] be want the things combined I'm just going to type it in because I know what I'm looking for I just don't know if it's on this laptop again the laptop is has been barely set up I tried to keep it as vanilla as possible um because I didn't want to have something that was like oh hey look at all this stuff that's already installed and you guys are like well he's he's got it already done I think it's interesting to watch M kind of fla and figure out what I'm trying to what I'm trying to find because you have to do the same thing find [Music] think okay there we go let's do and let's preview this mask here see what this look this should help so what I'm doing here I mean if you're like oh my God what did he just let's move all this junk out of the way here that's the old way so we're using this seg detector here so this is using ultral litics it's coming in we have to bring in the image as well and then we have a threshold a dial ation obviously making it bigger around the perimeter a crop factor and then a drop size and and then this down here is like what are you looking for and this is interesting all by the way is usually the best answer down here but let's look at how much better this mask may be it'll be here let's just run it and see what happens and there we go and it looks [Music] why pack to pack Ultra LS is finding it these thresholds are still the same let's change our threshold [Music] to how about the other direction [Music] seven I should be able to get a better answer out of this also type in the words like dress [Music] and this is why you always use the word all because I'm not really sure what these Wild Card Fields appear in a lot of different nodes and sometimes you can put in what you want to put in and sometimes you can't um I should get a much better mask than that out of this not very happy with that map in my opinion that is terrible oh what can I do to make it better anything better you know what it was better before I image scaled it so let's do that let's let's take it off the scale at least then it fills this side in [Music] let's do that let's get rid of all this here's my plan we're going to use segs here combine it to get the mask right but what I'm going to do is uh because this mask looks so much better here right I think that's kind of our answer is we're just going to use this mask this point forward uh to use for the in paint so where's my in paint uh I deleted that before [Music] I all [Music] that back to where we were so what did we do what was this whole train Rec about up here what I'm trying to do is I'm trying to create a mask using segs here right but I'm not getting a great result But ultimately the problem with this in my mind is the the tearing that occurs around the image by the way you can use the middle boss button to navigate without actually grabbing the image grabbing these things so the tearing because of the the resolution of the Mask is just not great U we could do uh another thing right now we're applying our mask here VA encoding right and we're growing The Mask let's not grow The Mask let's let's keep the mask at zero and see what and we're going to get we can see the pixels of the Mask in here so this is probably not going to look good either so cancel that what else can we do are there any other things that kind of help with this I don't think this helped us I thought we had a better solution beforehand so it just goes to show you that over complicating things doesn't necessarily mean you're going to get a better result I think this depth is not required here my mind this was not a positive I think this original latent here was a better latent this so much easier to look at but I mean that's that's not obviously the goal but that's nice when that happens and I think our our cany at 50% so we start at zero we start applying cany immediately so it holds the fingers in place uh but at halfway through which might still be a bit too far along we'll say back this up to 30% um that it's allowed to start changing stuff so if we watch this with the bar at the top there oh I have a problem node disconnected [Music] somewhere whenever whenever it flies too fast you know and oops work better but they'll get there right it's not a UI it's a it's a backend first so 30% through it has to hold the cany and then it's allowed to do whatever it wants to do so it's adding all the extra hairs or things but the fingers might be a little bit too free so let's take that and I actually kind of want to up the resolution of this Lis depth net here so it's the same 1152 make it a more accurate depth map and let's apply the cany a little bit later in the game I think point four cuz we're losing fingers cany is the only thing that's holding those in [Music] place try again smooth the Jazz do me feel like I'm in a dining club establishment yes dinner and a show there you go fingers are better is y fingers are better uh and we're getting more U more interesting places because the depth map is allowed some wandering around so if we were to go to the client now um how is this looking I think again I think this cany is still a bit too because I'm not you see this finger here is missing this finger here if you haven't worked with ballerinas before I I worked with her I think this is my second time it is next level because it's I obviously beautiful art form him but it is everything is so either right or wrong like if her foot is facing the wrong direction even slightly they consider to be this big atrocity and we're like hey that's a great picture and they're like oh my God don't show anybody that they're going to think I'm an idiot and we don't know better we show the world and they go wow you're a horrible photographer because you should have known that's a terrible pose I learned a lot about shooting dancers and wow I mean it's it's just so much you got to be careful about um that you don't even think about because we don't know you know what you don't know all right so I think this dress is looking kind of terrible so let's go and change it to oh let's just leave it get rid of the color mandate here I water uh wild butterfly we'll just let it do what it wants to do uh and we'll get rid of all this other mandate for this so and now let's do some so high detailed right uh let's throw the word fractal in there I like the word fractal volumetric another one I like if I can spell it volumetric that's kind of like fog in the in the distance who knows again the depth map is the one who takes control of this let's run it again and see if we're getting closer to again the idea is to feature dresses with the clients but yet have proper fing finger positions and let the dress kind of be what it wants to be and there we go so by lifting the Mandate of dress colors now it's actually able to shine and do what it wants so now I should be able to really kind of let this thing do what it wants to do so let's put our our goey back together so put the preview back up here for that put the preview for back up here just so that we have them there's one control net there's another control net got two control Nets being applied to [Music] this again I put this graph in the community area for all the sponsor Channel sponsors again thank you guys and girls so much for your support especially Tomac who is the only Channel God at the moment had a few over the last year or so which blows my mind you guys are amazing and then what the we'll put the prompt back over here all right so we reassembled this thing now we're in a situation where we can start cranking out images for our client right so a few other things that are interesting and this is not the turbo model we're still producing an image about every 10 seconds using sdxl and 20 steps and and we're getting a really nice image uh if we wanted to upscale this then I would look at using the SD ultimate upscaler that's pretty much the way I upscale everything I'm very happy with the way that works by the way if you like the show and and whatever please click the like button again we have people coming and going um and we have about two-thirds of the people have clicked a like button so again if you like it please please show your appreciation there all right so another thing here is this extra options button up here you can Auto CU and then now this is a preview we haven't actually saved I'm going to save this we actually saved a single image this entire time so I'm going to go here and I'm just going to click save image so we actually start saving these to the hard drive so I don't like miss them so if I hit this it's going to q1 render it and we have this set to increment so we're going to random one every time and as soon as it's done it should cue another one now the laptop's going to start sweating here in a second if we could look up the different styles of butterflies we could do Monarch we could do swallow tail we could do whatever uh and and have it go nuts but I think we're getting some really interesting stuff so someone said hey we want to do some butterfly themed dresses we're looking for ideas for Fabrics things along this lines can you give us some thoughts here we go we can start cranking them out yes we could probably do something to make sure that it's a same girl each and every time uh but we tried that with in painting and we really didn't like that too much we could try adding a control net to this my fear with control net is the control net is going to control the dress design too tightly right um I want to kind of be um be free to do whatever it wants to do I don't like the girl's hair okay we could also do what's called Regional prompting um I need to actually do a separate video on that where we can apply a prompt only to a region uh it will allow to change the rest of the image um but the rest of it want to be open uh you seen a video to fix hands using open pose control net to edit the number of fingers in Photoshop and use the image yeah the thing is though this is working really well like I'm getting I'm getting pretty much the proper fingers if we if we change this control net up a bit more this cany to say 70% I think that that would eliminate any of that why you guys hear that laptop I'm very happy with this thing by the way like I have to say that cuz people send me stuff all the time and m whatever but I'm really really happy with this one that MSI one I kind of had buyers remorse for the whole time although like when I go and I speak at different conferences I have to have something I can do inference on locally because I don't have access to my main machine and this is like a really good test for me to see if this was actually going to be able to pull its weight on these things um let me actually switch back to that you guys can see it there we go you can see it's actually pounding these things out pretty readily here once those tags get off the bottom of the screen there but that's on a laptop so yeah I'm very very happy with this thing moving right along so here's our dress designs all butterflies and whatever and I'm very happy with it we tried to implement uh another type of masking for imp painting and I think we ended up with a lot of tearing what we ended up with here looks pretty darn good again it's this is what what's this doing like think about what this is doing uh from the big picture standpoint let me go back to this ATM here um what this is doing from a big picture standpoint is showing how we can Implement uh a stable diffusion solution for a client that's a real a real world situ situation so clients like we want to be able to come up with different designs we are looking for ideas we spend months and months and months and this is a very true statement textile designers spend months trying to come up with ideas just brainstorming things that are different and unique and then they want to put them out and and show them to the world and try and get other people to to understand whether or not that they're going to be pertinent and if we can shortcut that process for them so that their designers are able to produce more winning images more quickly that means that they have better sales on the rack the stores you know have have better movement of product which means more things get stuff get sold more jobs get made and so on so I really like using a lot of this uh this technology in real world situations let see what we can do with it like there's so much to be done here I'd like to see control net pose input just to see if it'd be over constrained I think the control net works fine most of the time but I think it still has issues with fingers now there are like DW open pose which is better than regular open pose uses the Onyx system which means it's another 30 seconds uh in a loading time initially um that's not a big deal but I don't know that it adds that much better control of hands uh especially when there's such a minor part of the of this if they were hands that were you know larg and in charge in the scene or holding something that's more important I think it's a bigger issue uh but I didn't want to kind of go down the pose band the pose bandwagon or the pose Direction because pose is very cool but pose opens a whole lot of other problems for us to deal with and I think that'd be a great separate live stream or a great separate video to cover because you're not wrong I think it's a great product um but the the idea behind the live stream here is to kind of show how we could use this and say let's be constrained let's use just a few tools what few tools can we use are there better ways to do this in 3 months this will probably be old news and there'll be a brand way to do it that'll be a lot better but right now I'm able to actually open the doors and I've done this this is a very true statement we have gone to manufacturers and said let us help you with your ideation workflow and we've implemented stable diffusion in that workflow to say Let's Help designers brainstorm quickly especially when turbo came along you know to be able to go and manipulate through these things very quickly it's been fantastic now with 0 1 2 3 and some other ones like this we can actually start to rotate the product and do things like that that's cool but it has no practical application that I've found yet now again it's very very new the trouble we you kind of got to keep in mind is that when you're trying to sell a product the the product the picture of the product that's on the box needs to actually be the product that's called Truth and advertising so uh we always you always hear people like talk about when you photograph ice cream just use mashed potatoes because they won't melt and you have more time to take the picture that's very true unless you're selling ice cream if you are selling the ice cream you're taking a picture of you cannot use mash potatoes you have to actually use the ice cream you're selling that's truth and advertising the same thing applies to all the images are producing here we cannot sell this dress and say this is the dress you're buying because this is not a photo of that dress right it is an AI representation of that dress and that's not truth and advertising and if someone gets this dress and said that's not what I ordered you have to refund it and you're going to go out of business cuz that's not not going to work very well so we have to walk that fine line you know everyone who's I think in this channel here you're probably looking at this is like how do you wrap a business around this how do you do things around it understanding the ramifications of where this fits in the workflow for somebody and what you're trying to sell we are not selling a picture of this girl to anybody right what we're doing is we're using this to try and help people design a better dress or a theme like you were going to go to a ren fair and Sam I'm trying to come up with some really creative armor ideas let's use this to create really creative armor ideas we could this dress as armor maybe I'll do that I don't know you can do that you download this workflow and you play with it you come up with some armor dresses and that's maybe an idea that somebody's going to say I want to make that out of leather I'm going to make that out of chain mail or SK mail and then you're going to sell it you've given them ideas you give them a place to go to to get those ideas you like you become your the idea person and then that's one of the really great things I love about this I don't use it as a finished product very often I use it more as an ideation process but that's just one aspect of this whole thing there's so many ways you can take this and again it's not even it's like a little over a year old I work for stability now is it 14 months now what a wild ride it's been and I I can't wait to see where we're going you know in the future all the stuff that we can do is just really really wild but um anyway that's the live stream for today you know I really appreciate you guys hanging out with me today appreciate the the two new people who joined today and we passed 300 members appreciate gigabyte for sending the 17x laptop over and again we're using that for inference on this so we'll be using that for uh inference U from now on and I'm going to try and move all the models and stuff over here that I'm using and again I don't use a lot so it's not a big deal I use like maybe two or three luras like plenty but let's explore them more I like exploring the Lura I'm going to uh produce the next video I'm going to do us more on how to really bend a Laura to make it do what you want to do CU we apply luras with a strength but there's a lot more you can do the Laura you can interrupt how it's interfacing with the unet down below all kinds of cool stuff in there so we'll do that in the future I could try and get that video out in the next couple days or so but appreciate everybody hanging out with me today again thank you very much I will put the patron screen up at the end but again I really appreciate you guys and please make sure you check the community area in YouTube here because that's how I have to communicate with you YouTube doesn't have a great method like patreon does and I really don't want to move over to patreon I kind of want to keep things here on YouTube because then that way you guys have much easier access to the videos than trying to have a VI Vimeo login and all that kind of crap let's just keep it here but make sure you check that area right even if you're not a member I do post things there for non-members from time in time but U if you're a member you're going to get this workflow and a lot of other stuffff like got coming out I I try and post something every week guys so oh and there's new embeddings coming out as well may guys embeddings that you can use with the IP adapter to add manipulation at an artistic level to some of the imagery so no matter what model you're using or Lura you end up with kind of a uh kind of a different look which is my goal I'm trying to make these more artistic the whole photogenic idea it's cool but it it's not really my core interest it might be yours but it's not mine but again these tools are for your interpretation all I'm doing is trying to teach how to use them everybody take care stay safe and I'll catch you all next [Music] time [Music] [Applause] [Music] a [Music] n

Channel: Scott Detweiler

Views: 12,550

Keywords: ai art, creativity, ai art tutorial, ai generated art, ai art generator, ai, ai created art, photobashing, stable diffusion, ai painting, #ai 画像 生成, comfyUI tutorial, comfy tutorial, stable diffusion tutorial, comfyui

Id: Pog7CoRKQRw

Length: 95min 11sec (5711 seconds)

Published: Sun Dec 31 2023